An AI readiness assessment is a check, made before an AI project starts, of whether an organization has what that project needs: data it can use, the skills to build and run it, rules for what the system may do, and a measure of success. TDWI’s published assessment gives data one of its five categories. This post goes deep on that one.
Key takeaways
- In a Gartner survey of 1,203 data management leaders (July 2024, published February 26, 2025), 63% of organizations either did not have, or were unsure whether they had, the right data management practices for AI.
- In the same release Gartner predicted that through 2026 organizations would abandon 60% of AI projects that lack AI-ready data, and said AI-ready data has to be representative of the use case it serves.
- TDWI’s AI Readiness Assessment (April 8, 2024) has about 75 questions across five categories: organizational, data, skills, operational and governance readiness.
- For high-risk AI systems, Article 10 of the EU AI Act (Regulation (EU) 2024/1689) requires the data sets used to train, validate and test the system to be governed and managed in ways that suit its intended purpose. After the July 2026 amendment (Regulation (EU) 2026/1744), that requirement applies from December 2, 2027 for systems listed in Annex III and from August 2, 2028 for those under Annex I.
- UNESCO’s Readiness Assessment Methodology measures countries, not organizations; UNESCO says more than 70 countries have implemented it or are implementing it.
What is an AI readiness assessment?
It answers a go or no-go question for a specific use. “Are we ready for AI?” is too broad to decide a project on; “Is our claims data ready for a model that triages new claims?” isn’t. The assessment looks at the use case, the data it needs, the people and platforms that will run it, and the rules it must follow, and ends in a decision with conditions attached.
Published frameworks differ in where they put the weight. TDWI’s online assessment gives data readiness one of its five categories. UNESCO’s Readiness Assessment Methodology works at the level of a country’s laws and institutions, so it is useful as a contrast but not as a template for a company. In our view, an organization-level AI readiness framework needs most of its depth in data, the part a questionnaire can test against real records.
Readiness or maturity?
The two get used as if they were the same, and they answer different questions. AI readiness asks whether you can start this project now. AI maturity describes how well an organization already uses AI across many projects, in stages; our post on the AI maturity model covers those stages.
Data maturity sits underneath both. A data maturity model scores the whole organization’s data management area by area, and those scores are a shortcut for a readiness check.
The data areas that decide AI readiness
Gartner’s February 2025 release sets out five steps toward AI-ready data. In our words: align data to the use case, work out the governance requirements, move metadata from a passive record to an active one, prepare the pipelines for training and production, and keep testing and monitoring the data once the model uses it. It also warns that organizations with basic or manual metadata practices will struggle to make their data AI-ready.
Mapped to the data management areas we assess, those steps land in a handful of places:
- Data quality matters because errors in the training data become errors in the model’s output. The data quality dimensions that matter most depend on the use: for example, timeliness for a forecasting model, or accurate labels for a classifier.
- Metadata and lineage supply the definitions, the owners and the path each field took from its source. Without them nobody can say whether the training data represents the cases the model will see; metadata management is the area to check.
- Security and privacy decide whether personal data may be used for this new purpose, and who can see what the model sees.
- Governance settles who approves a use case, who owns the model’s data once it is live, and who can stop it. AI data governance covers this through DCAM’s analytics management component.
- Documents are the data for a generative AI assistant that answers from your own files. Keeping their versions current and their access rights correct belongs to the DMBOK area of document and content management.
For a high-risk system under the EU AI Act, Article 10 turns several of these into obligations once the high-risk requirements apply (from December 2, 2027 or August 2, 2028, depending on the kind of system). The data governance practices it lists include the origin of the data and, for personal data, the purpose it was first collected for; preparation steps such as labeling and cleaning; an assessment of whether enough suitable data is available; an examination for possible biases; and the identification of data gaps. Article 10 also asks that the data sets be “relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose.”
AI readiness checklist
Score the checklist for one use case, with the owner of the data and the person who will build the model in the same meeting. Each row is a statement; mark it true, partly true or false, and write down the evidence.
| Area | Ready when | Warning sign |
|---|---|---|
| Use case | The decision the AI will support is written in one sentence, with the measure that will show it worked | The use case is “use AI on our data” |
| Data scope | Every data set the use case needs is listed with its source system and a named owner | A data set has no owner, or nobody knows where it comes from |
| Representativeness | The available history covers the cases the model will meet in production, including the rare ones | The training data comes from one region, product or period only |
| Data quality | The fields the model depends on have been profiled, with error rates known | Quality is described as “pretty good” with no figures |
| Labels | Where the model learns from past outcomes, the outcome field is recorded consistently and someone has checked a sample | Outcome codes changed meaning partway through the history |
| Metadata | Each field used has a business definition and an owner who answers questions about it | The team learns what a field means by reading the code that fills it |
| Lineage | The path from source to training set can be traced and rerun | The training extract was built once, by hand, by someone who has left |
| Privacy and legal basis | The use of any personal data in the model has been checked against the purpose it was collected for | Nobody has asked the privacy officer |
| Access and security | The model, its prompts and its outputs can only reach data the user is allowed to see | A generative assistant is pointed at a shared drive with no permission review |
| Documents (generative AI) | Source documents have owners, and superseded versions are filtered out of the index | Three versions of the same policy sit side by side in the index |
| Pipelines | The data can reach the model in production on the schedule the use case needs | The pilot ran on a spreadsheet export that production can’t repeat |
| Monitoring | Someone owns checking the input data and the outputs after go-live, with a threshold that triggers a review | Monitoring is planned “after launch” |
| Decision rights | One named person can approve the use case, and one can switch it off | Approval is “the steering committee” with no date |
The checklist has no bias row of its own because the representativeness and labels rows carry it: a data set that misses a group of customers, or labels them inconsistently, is where bias starts. Where the system could count as high-risk, add the bias examination Article 10 describes as its own row.
How to run one in five steps
- Pick one use case. Write the decision it supports and how you will measure success. A readiness check for “AI in general” has nothing to test.
- List the data. Name every data set, its source system and its owner, and invite those owners.
- Read what you already know. If the organization has run a data maturity assessment, pull the scores for Data Quality, Metadata Management, Data Security and Data Governance before the meeting.
- Score the checklist and test a sample. Work through the table above, then profile a sample of the actual data the model would use, so the quality and representativeness rows rest on figures.
- Decide, with conditions and dates. Record the decision as go, conditional go or not yet, using the rules below, and set the date for the next check.
Go, conditional go, not yet
The decision rules are ours. Agree on them with the sponsor before scoring, so the result can’t be argued afterward.
| Decision | When it applies | What happens next |
|---|---|---|
| Go | Every row is true, or partly true with a fix that doesn’t block a pilot | Build, with the monitoring row in the plan from day one |
| Conditional go | The governance rows (use case, data scope, privacy, decision rights) are true or partly true with a non-blocking fix, and at least one data row is false or has a blocking gap | A time-boxed pilot on the data that passes, while named owners fix the rest by a date |
| Not yet | Any governance row is false, or partly true with a fix that blocks a pilot | Fix those first; they are governance work for the data owners and the sponsor |
For illustration: a manufacturer wants an assistant that answers its service technicians’ questions from equipment manuals. The governance rows (use case, data scope, privacy, decision rights) are all true. The documents row is false, because the shared drive holds each manual’s past revisions next to the current one, and the access row is partly true. That makes it a conditional go: a pilot on one product line whose manuals the documentation team has cleaned, while the same team retires superseded revisions across the rest.
In our view, a string of “not yet” decisions across several use cases is a sign that the gaps sit in data management rather than in any one project. That is the point to put AI readiness into the data strategy, and to decide how the organization governs AI as a whole, which our post on the AI governance framework covers.
Scoring the organization around the use case
The checklist stops at one use case. It doesn’t ask about the strategy behind the project, the platform it runs on, who leads the AI work or whether staff will use the result, because those belong to the organization rather than to one project. Data Maturity Pro’s free AI readiness assessment scores them in eight areas, with Data Readiness for AI as one of them, and places each area on a level from Initial to Optimized. Its Quick version asks two questions per area, 16 in all, in about 10 minutes, and the Detailed version asks 56.
Run it before the second or third use case, when the same gaps start to come back, and keep using the checklist above for each project. For the governance side on its own, our AI governance maturity model has a rubric you can score against evidence.
DAMA-DMBOK and DMBOK are trademarks of DAMA International. DCAM is a trademark of the EDM Council (now the EDM Association). Level Up Your Data is not affiliated with or endorsed by DAMA International or the EDM Council.
Frequently asked questions
What does an AI readiness assessment include?
TDWI’s version asks about 75 questions in five categories, data among them. Ours is a 13-row checklist for one use case: four governance rows (use case, data scope, privacy, decision rights) and nine on the data itself, each marked true, partly true or false with its evidence.
How do you measure AI readiness?
Against a single use case, with evidence for each item rather than opinions. An online questionnaire gives a score per category, and a data maturity score per area tells you where to look. The final measure is a decision: go, conditional go or not yet.
Is data maturity required for AI?
In our view, not across the board: a project needs mature practice in the data it touches, not in every area of data management. So score only the areas this project relies on, starting with the ones its data sets depend on.
How long does an AI readiness assessment take?
It depends on the scope. TDWI’s online assessment returns scores once you submit it; its page gives no completion time. In our view, the data checks for one use case need one working session with the data owners plus the time to profile a sample of the data.
Summary
AI readiness is decided use case by use case, and the data rows are where the conditions come from. Pick the use case you most want to fund this year and score the checklist against its data, with its owners in the room.